def retrace_stream_plan(store, domain, goal_expression): # TODO: retrace the stream plan used if store.best_plan is None: return None assert not domain.axioms from pddlstream.algorithms.downward import plan_preimage print(goal_expression) plan_preimage(store.best_plan, goal_expression) raise NotImplementedError()
def recover_negative_axioms(real_task, opt_task, axiom_plans, action_plan, negative_from_name): action_plan = reinstantiate_action_instances(opt_task, action_plan) simplify_conditional_effects(opt_task, action_plan, negative_from_name) axiom_plans = list(map(reinstantiate_axiom_instances, axiom_plans)) axioms_from_name = get_derived_predicates(opt_task.axioms) # TODO: could instead just accumulate difference between real and opt opt_task.init = set(opt_task.init) real_states = [set(real_task.init)] preimage_plan = [] for axiom_plan, action_instance in safe_zip(axiom_plans, action_plan): preimage = list(plan_preimage(axiom_plan + [action_instance], [])) assert conditions_hold( opt_task.init, (l for l in preimage if (l.predicate not in axioms_from_name and ( l.predicate not in negative_from_name)))) # TODO: only add derived facts and negative facts to fluent state to make normalizing easier new_axiom_plan = extract_axiom_plan(opt_task, preimage, negative_from_name, static_state=opt_task.init) #static_state=real_states[-1]) assert new_axiom_plan is not None preimage_plan.extend(new_axiom_plan + axiom_plan + [action_instance]) if action_instance.name != GOAL_NAME: apply_action(opt_task.init, action_instance) real_states.append(set(real_states[-1])) apply_action(real_states[-1], action_instance) return real_states, preimage_plan
def recover_negative_axioms(real_task, opt_task, axiom_plans, action_plan, negative_from_name): action_plan = reinstantiate_action_instances( opt_task, action_plan, negative_from_name=negative_from_name) # https://github.com/caelan/pddlstream/commit/18b303e19bbab9f8e0016fbb2656f461067e1e94#diff-55454a85485551f9139e20a446b56a83L53 #simplify_conditional_effects(opt_task, action_plan, negative_from_name) axiom_plans = list(map(reinstantiate_axiom_instances, axiom_plans)) axioms_from_name = get_derived_predicates(opt_task.axioms) # TODO: could instead just accumulate difference between real and opt opt_task.init = set(opt_task.init) real_states = [set(real_task.init)] preimage_plan = [] for axiom_plan, action_instance in safe_zip(axiom_plans, action_plan): preimage = [ l for l in plan_preimage(axiom_plan + [action_instance]) if (l.predicate in axioms_from_name) ] #assert conditions_hold(opt_task.init, conditions) # TODO: only add derived facts and negative facts to fluent state to make normalizing easier negative_axiom_plan = extract_axiom_plan(opt_task, preimage, negative_from_name, static_state=opt_task.init) #static_state=real_states[-1]) assert negative_axiom_plan is not None preimage_plan.extend(negative_axiom_plan + axiom_plan + [action_instance]) if action_instance.name != GOAL_NAME: apply_action(opt_task.init, action_instance) real_states.append(set(real_states[-1])) apply_action(real_states[-1], action_instance) return real_states, preimage_plan
def recover_negative_axioms(real_task, opt_task, axiom_plans, action_plan, negative_from_name): action_plan = reinstantiate_action_instances(opt_task, action_plan) axiom_plans = list(map(reinstantiate_axiom_instances, axiom_plans)) axioms_from_name = get_derived_predicates(opt_task.axioms) # TODO: could instead just accumulate difference between real and opt opt_task.init = set(opt_task.init) real_states = [set(real_task.init)] preimage_plan = [] for axiom_plan, action_instance in safe_zip(axiom_plans, action_plan): for literal in action_instance.precondition: # TODO: check conditional effects if literal.predicate in negative_from_name: raise NotImplementedError( 'Negated predicates not currently supported within actions: {}' .format(literal.predicate)) simplify_conditional_effects(real_states[-1], opt_task.init, action_instance, axioms_from_name) preimage = list(plan_preimage(axiom_plan + [action_instance], [])) assert conditions_hold( opt_task.init, (l for l in preimage if l.predicate not in axioms_from_name)) new_axiom_plan = extract_axiom_plan(opt_task, preimage, negative_from_name, static_state=real_states[-1]) assert new_axiom_plan is not None preimage_plan.extend(new_axiom_plan + axiom_plan + [action_instance]) if action_instance.name != GOAL_NAME: apply_action(opt_task.init, action_instance) real_states.append(set(real_states[-1])) apply_action(real_states[-1], action_instance) return real_states, preimage_plan
def recover_stream_plan(evaluations, current_plan, opt_evaluations, goal_expression, domain, node_from_atom, action_plan, axiom_plans, negative): # Universally quantified conditions are converted into negative axioms # Existentially quantified conditions are made additional preconditions # Universally quantified effects are instantiated by doing the cartesian produce of types (slow) # Added effects cancel out removed effects # TODO: node_from_atom is a subset of opt_evaluations (only missing functions) real_task = task_from_domain_problem( domain, get_problem(evaluations, goal_expression, domain)) opt_task = task_from_domain_problem( domain, get_problem(opt_evaluations, goal_expression, domain)) negative_from_name = get_negative_predicates(negative) real_states, combined_plan = recover_negative_axioms( real_task, opt_task, axiom_plans, action_plan, negative_from_name) function_plan = compute_function_plan(opt_evaluations, action_plan) full_preimage = plan_preimage(combined_plan, []) stream_preimage = set(full_preimage) - real_states[0] negative_preimage = set( filter(lambda a: a.predicate in negative_from_name, stream_preimage)) function_plan.update( convert_negative(negative_preimage, negative_from_name, full_preimage, real_states)) positive_preimage = stream_preimage - negative_preimage step_from_fact = { fact_from_fd(l): full_preimage[l] for l in positive_preimage if not l.negated } target_facts = { fact for fact in step_from_fact.keys() if get_prefix(fact) != EQ } #stream_plan = reschedule_stream_plan(evaluations, target_facts, domain, stream_results) # visualize_constraints(map(fact_from_fd, target_facts)) stream_plan = [] for result in current_plan: if isinstance(result.external, Function) or (result.external in negative): function_plan.add( result) # Prevents these results from being pruned else: stream_plan.append(result) curr_evaluations = evaluations_from_stream_plan(evaluations, stream_plan, max_effort=None) extraction_facts = target_facts - set( map(fact_from_evaluation, curr_evaluations)) extract_stream_plan(node_from_atom, extraction_facts, stream_plan) stream_plan = postprocess_stream_plan(evaluations, domain, stream_plan, target_facts) stream_plan = convert_fluent_streams(stream_plan, real_states, action_plan, step_from_fact, node_from_atom) return stream_plan + list(function_plan)
def recover_axioms_plans(instantiated, action_instances): task = instantiated.task derived_predicates = get_derived_predicates(task.axioms) state = set(task.init) axiom_plans = [] for action_instance in action_instances + [get_goal_instance(task.goal)]: # TODO: apply all axiom_instances unaffected by negative conditions preimage = list(plan_preimage([action_instance], [])) axiom_instances = filter( lambda ax: all(l.predicate in derived_predicates or literal_holds( state, l) for l in ax.condition), instantiated.axioms) # Only instantiate if preimage has goal axiom_plan = extraction_helper(state, axiom_instances, preimage) assert axiom_plan is not None axiom_plans.append(axiom_plan) apply_action(state, action_instance) return axiom_plans
def recover_opt_stream_plan(evaluations, action_plan, task): plan_instances = get_action_instances(task, action_plan) + [get_goal_instance(task.goal)] replace_derived(task, set(), plan_instances) preimage = filter(lambda a: not a.negated, plan_preimage(plan_instances, [])) opt_stream_plan = [] opt_from_obj = {} for stream_result in extract_order(evaluations, preimage): input_objects = tuple(opt_from_obj.get(o, o) for o in stream_result.instance.input_objects) instance = stream_result.instance.external.get_instance(input_objects) #instance.opt_index = stream_result.instance.opt_index #assert(instance.opt_index == 0) instance.opt_index = 0 #assert(not instance.disabled) instance.disabled = False opt_results = instance.next_optimistic() if not opt_results: continue opt_result = opt_results[0] opt_stream_plan.append(opt_result) for obj, opt in zip(stream_result.output_objects, opt_result.output_objects): opt_from_obj[obj] = opt return opt_stream_plan, opt_from_obj
def recover_stream_plan(evaluations, current_plan, opt_evaluations, goal_expression, domain, node_from_atom, action_plan, axiom_plans, negative, replan_step): # Universally quantified conditions are converted into negative axioms # Existentially quantified conditions are made additional preconditions # Universally quantified effects are instantiated by doing the cartesian produce of types (slow) # Added effects cancel out removed effects # TODO: node_from_atom is a subset of opt_evaluations (only missing functions) real_task = task_from_domain_problem( domain, get_problem(evaluations, goal_expression, domain)) opt_task = task_from_domain_problem( domain, get_problem(opt_evaluations, goal_expression, domain)) negative_from_name = { external.blocked_predicate: external for external in negative if external.is_negated() } real_states, full_plan = recover_negative_axioms(real_task, opt_task, axiom_plans, action_plan, negative_from_name) function_plan = compute_function_plan(opt_evaluations, action_plan) full_preimage = plan_preimage(full_plan, []) # Does not contain the stream preimage! negative_preimage = set( filter(lambda a: a.predicate in negative_from_name, full_preimage)) negative_plan = convert_negative(negative_preimage, negative_from_name, full_preimage, real_states) function_plan.update(negative_plan) # TODO: OrderedDict for these plans # TODO: this assumes that actions do not negate preimage goals positive_preimage = { l for l in (set(full_preimage) - real_states[0] - negative_preimage) if not l.negated } steps_from_fact = { fact_from_fd(l): full_preimage[l] for l in positive_preimage } last_from_fact = { fact: min(steps) for fact, steps in steps_from_fact.items() if get_prefix(fact) != EQ } #stream_plan = reschedule_stream_plan(evaluations, target_facts, domain, stream_results) # visualize_constraints(map(fact_from_fd, target_facts)) for result, step in function_plan.items(): for fact in result.get_domain(): last_from_fact[fact] = min(step, last_from_fact.get(fact, INF)) # TODO: get_steps_from_stream stream_plan = [] last_from_stream = dict(function_plan) for result in current_plan: # + negative_plan? # TODO: actually compute when these are needed + dependencies last_from_stream[result] = 0 if isinstance(result.external, Function) or (result.external in negative): if len(action_plan) != replan_step: raise NotImplementedError( ) # TODO: deferring negated optimizers # Prevents these results from being pruned function_plan[result] = replan_step else: stream_plan.append(result) curr_evaluations = evaluations_from_stream_plan(evaluations, stream_plan, max_effort=None) extraction_facts = set(last_from_fact) - set( map(fact_from_evaluation, curr_evaluations)) extract_stream_plan(node_from_atom, extraction_facts, stream_plan) # Recomputing due to postprocess_stream_plan stream_plan = postprocess_stream_plan(evaluations, domain, stream_plan, last_from_fact) node_from_atom = get_achieving_streams(evaluations, stream_plan, max_effort=None) fact_sequence = [set(result.get_domain()) for result in stream_plan] + [extraction_facts] for facts in reversed(fact_sequence): # Bellman ford for fact in facts: # could flatten instead result = node_from_atom[fact].result if result is None: continue step = last_from_fact[fact] if result.is_deferrable() else 0 last_from_stream[result] = min(step, last_from_stream.get(result, INF)) for domain_fact in result.instance.get_domain(): last_from_fact[domain_fact] = min( last_from_stream[result], last_from_fact.get(domain_fact, INF)) stream_plan.extend(function_plan) # Useful to recover the correct DAG partial_orders = set() for child in stream_plan: # TODO: account for fluent objects for fact in child.get_domain(): parent = node_from_atom[fact].result if parent is not None: partial_orders.add((parent, child)) #stream_plan = topological_sort(stream_plan, partial_orders) bound_objects = set() for result in stream_plan: if (last_from_stream[result] == 0) or not result.is_deferrable(bound_objects=bound_objects): for ancestor in get_ancestors(result, partial_orders) | {result}: # TODO: this might change descendants of ancestor. Perform in a while loop. last_from_stream[ancestor] = 0 if isinstance(ancestor, StreamResult): bound_objects.update(out for out in ancestor.output_objects if out.is_unique()) #local_plan = [] # TODO: not sure what this was for #for fact, step in sorted(last_from_fact.items(), key=lambda pair: pair[1]): # Earliest to latest # print(step, fact) # extract_stream_plan(node_from_atom, [fact], local_plan, last_from_fact, last_from_stream) # Each stream has an earliest evaluation time # When computing the latest, use 0 if something isn't deferred # Evaluate each stream as soon as possible # Option to defer streams after a point in time? # TODO: action costs for streams that encode uncertainty state = set(real_task.init) remaining_results = list(stream_plan) first_from_stream = {} #assert 1 <= replan_step # Plan could be empty for step, instance in enumerate(action_plan): for result in list(remaining_results): # TODO: could do this more efficiently if need be domain = result.get_domain() + get_fluent_domain(result) if conditions_hold(state, map(fd_from_fact, domain)): remaining_results.remove(result) certified = { fact for fact in result.get_certified() if get_prefix(fact) != EQ } state.update(map(fd_from_fact, certified)) if step != 0: first_from_stream[result] = step # TODO: assumes no fluent axiom domain conditions apply_action(state, instance) #assert not remaining_results # Not true if retrace if first_from_stream: replan_step = min(replan_step, *first_from_stream.values()) eager_plan = [] results_from_step = defaultdict(list) for result in stream_plan: earliest_step = first_from_stream.get(result, 0) latest_step = last_from_stream.get(result, 0) assert earliest_step <= latest_step defer = replan_step <= latest_step if not defer: eager_plan.append(result) # We only perform a deferred evaluation if it has all deferred dependencies # TODO: make a flag that also allows dependencies to be deferred future = (earliest_step != 0) or defer if future: future_step = latest_step if defer else earliest_step results_from_step[future_step].append(result) # TODO: some sort of obj side-effect bug that requires obj_from_pddl to be applied last (likely due to fluent streams) eager_plan = convert_fluent_streams(eager_plan, real_states, action_plan, steps_from_fact, node_from_atom) combined_plan = [] for step, action in enumerate(action_plan): combined_plan.extend(result.get_action() for result in results_from_step[step]) combined_plan.append( transform_action_args(pddl_from_instance(action), obj_from_pddl)) # TODO: the returned facts have the same side-effect bug as above # TODO: annotate when each preimage fact is used preimage_facts = { fact_from_fd(l) for l in full_preimage if (l.predicate != EQ) and not l.negated } for negative_result in negative_plan: # TODO: function_plan preimage_facts.update(negative_result.get_certified()) for result in eager_plan: preimage_facts.update(result.get_domain()) # Might not be able to regenerate facts involving the outputs of streams preimage_facts.update( result.get_certified()) # Some facts might not be in the preimage # TODO: record streams and axioms return eager_plan, OptPlan(combined_plan, preimage_facts)
def recover_stream_plan(evaluations, current_plan, opt_evaluations, goal_expression, domain, node_from_atom, action_plan, axiom_plans, negative, replan_step): # Universally quantified conditions are converted into negative axioms # Existentially quantified conditions are made additional preconditions # Universally quantified effects are instantiated by doing the cartesian produce of types (slow) # Added effects cancel out removed effects # TODO: node_from_atom is a subset of opt_evaluations (only missing functions) real_task = task_from_domain_problem( domain, get_problem(evaluations, goal_expression, domain)) opt_task = task_from_domain_problem( domain, get_problem(opt_evaluations, goal_expression, domain)) negative_from_name = { external.blocked_predicate: external for external in negative if external.is_negated() } real_states, combined_plan = recover_negative_axioms( real_task, opt_task, axiom_plans, action_plan, negative_from_name) function_plan = compute_function_plan(opt_evaluations, action_plan) # TODO: record the supporting facts full_preimage = plan_preimage(combined_plan, []) stream_preimage = set(full_preimage) - real_states[0] negative_preimage = set( filter(lambda a: a.predicate in negative_from_name, stream_preimage)) function_plan.update( convert_negative(negative_preimage, negative_from_name, full_preimage, real_states)) positive_preimage = stream_preimage - negative_preimage steps_from_fact = { fact_from_fd(l): full_preimage[l] for l in positive_preimage if not l.negated } target_facts = { fact for fact in steps_from_fact.keys() if get_prefix(fact) != EQ } #stream_plan = reschedule_stream_plan(evaluations, target_facts, domain, stream_results) # visualize_constraints(map(fact_from_fd, target_facts)) # TODO: get_steps_from_stream stream_plan = [] step_from_stream = {} for result in current_plan: # TODO: actually compute when these are needed + dependencies step_from_stream[result] = 0 if isinstance(result.external, Function) or (result.external in negative): function_plan.add( result) # Prevents these results from being pruned else: stream_plan.append(result) curr_evaluations = evaluations_from_stream_plan(evaluations, stream_plan, max_effort=None) extraction_facts = target_facts - set( map(fact_from_evaluation, curr_evaluations)) step_from_fact = { fact: min(steps_from_fact[fact]) for fact in extraction_facts } extract_stream_plan(node_from_atom, extraction_facts, stream_plan, step_from_fact, step_from_stream) stream_plan = postprocess_stream_plan(evaluations, domain, stream_plan, target_facts) eager_plan = [] actions_from_step = {} for result in (stream_plan + list(function_plan)): if (result.opt_index != 0) or (step_from_stream.get(result, 0) < replan_step): eager_plan.append(result) else: actions_from_step.setdefault(step_from_stream[result], []).append(result.get_action()) eager_plan = convert_fluent_streams(eager_plan, real_states, action_plan, steps_from_fact, node_from_atom) # print(action_plan) # # TODO: propagate this forward in the future # start_from_stream = {} # for result in eager_plan: # stuff = list(map(fd_from_fact, get_fluent_domain(result))) # index = len(real_states) # for i, state in enumerate(real_states): # if conditions_hold(state, stuff): # start_from_stream[result] = i # index = i # break # #else: # #start_from_stream[result] = len(real_states) # print(index, result) # TODO: some sort of obj side-effect bug that requires obj_from_pddl to be applied last (likely due to fluent streams) #action_plan = transform_plan_args(map(pddl_from_instance, action_instances), obj_from_pddl) for step, action in enumerate(action_plan): actions_from_step.setdefault(step, []).append( transform_action_args(pddl_from_instance(action), obj_from_pddl)) action_plan = list( flatten(actions_from_step[step] for step in sorted(actions_from_step))) return eager_plan, action_plan
def recover_stream_plan(evaluations, goal_expression, domain, stream_results, action_plan, negative, unit_costs): import pddl import instantiate # Universally quantified conditions are converted into negative axioms # Existentially quantified conditions are made additional preconditions # Universally quantified effects are instantiated by doing the cartesian produce of types (slow) # Added effects cancel out removed effects real_task = task_from_domain_problem(domain, get_problem(evaluations, goal_expression, domain, unit_costs)) node_from_atom = get_achieving_streams(evaluations, stream_results) opt_evaluations = apply_streams(evaluations, stream_results) opt_task = task_from_domain_problem(domain, get_problem(opt_evaluations, goal_expression, domain, unit_costs)) function_assignments = {fact.fluent: fact.expression for fact in opt_task.init # init_facts if isinstance(fact, pddl.f_expression.FunctionAssignment)} type_to_objects = instantiate.get_objects_by_type(opt_task.objects, opt_task.types) results_from_head = get_results_from_head(opt_evaluations) action_instances = instantiate_actions(opt_task, type_to_objects, function_assignments, action_plan) negative_from_name = get_negative_predicates(negative) axioms_from_name = get_derived_predicates(opt_task.axioms) opt_task.init = set(opt_task.init) real_states = [set(real_task.init)] # TODO: had old way of doing this (~July 2018) preimage_plan = [] function_plan = set() for layer in action_instances: for pair, action_instance in layer: axiom_plan = extract_axiom_plan(opt_task, action_instance, negative_from_name, static_state=real_states[-1]) if axiom_plan is None: continue simplify_conditional_effects(real_states[-1], opt_task.init, action_instance, axioms_from_name) preimage_plan.extend(axiom_plan + [action_instance]) apply_action(opt_task.init, action_instance) real_states.append(set(real_states[-1])) apply_action(real_states[-1], action_instance) if not unit_costs and (pair is not None): function_result = extract_function_results(results_from_head, *pair) if function_result is not None: function_plan.add(function_result) break else: raise RuntimeError('No action instances are applicable') # TODO: could instead just accumulate difference between real and opt full_preimage = plan_preimage(preimage_plan, []) stream_preimage = set(full_preimage) - real_states[0] negative_preimage = set(filter(lambda a: a.predicate in negative_from_name, stream_preimage)) positive_preimage = stream_preimage - negative_preimage function_plan.update(convert_negative(negative_preimage, negative_from_name, full_preimage, real_states)) step_from_fact = {fact_from_fd(l): full_preimage[l] for l in positive_preimage if not l.negated} target_facts = list(step_from_fact.keys()) #stream_plan = reschedule_stream_plan(evaluations, target_facts, domain, stream_results) stream_plan = [] extract_stream_plan(node_from_atom, target_facts, stream_plan) stream_plan = prune_stream_plan(evaluations, stream_plan, target_facts) stream_plan = convert_fluent_streams(stream_plan, real_states, step_from_fact, node_from_atom) # visualize_constraints(map(fact_from_fd, stream_preimage)) if DO_RESCHEDULE: # TODO: detect this based on unique or not # TODO: maybe test if partial order between two ways of achieving facts, if not prune new_stream_plan = reschedule_stream_plan(evaluations, target_facts, domain, stream_plan) if new_stream_plan is not None: stream_plan = new_stream_plan return stream_plan + list(function_plan)